Alibaba Unveils Qwen3.5 AI Model, Outperforms GPT 5.2 & Claude Opus 4.5

Alibaba Group Holding Ltd. has introduced a new artificial intelligence model, Qwen3.5, claiming it can outperform GPT-5.2 and Anthropic’s Claude Opus 4.5 on select tasks.

The model is being released under an open-source license on Hugging Face, giving developers access to its capabilities and customization tools.

By default, Qwen3.5 supports prompts up to 262,144 tokens, with developers able to expand that capacity significantly. The system can process text in more than 210 languages and dialects alongside images such as charts and visualizations. It uses a mixture-of-experts architecture with 397 billion parameters overall, though only 17 billion are activated per prompt, improving hardware efficiency.

Alibaba has integrated optimisations to reduce computing demands, including a hybrid attention system that blends quadratic and linear attention to lower memory usage.

The model also incorporates a gated delta network, combining gating techniques and parameter-update methods to streamline training and inference efficiency. Research from Nvidia Corp. previously suggested such combinations can cut hardware requirements for large models.

In benchmark testing across more than 30 evaluations, Qwen3.5 outperformed rivals on instruction-following tests and certain reasoning benchmarks, while showing strong multimodal performance in visual reasoning and coding tasks.

Last year, the Chinese e-commerce giant launched Qwen3-Max, its most powerful AI model with over 1 trillion parameters, at its annual conference on Wednesday, marking the company’s largest-ever leap in AI development.